DeepSeek-R1 · GitHub Models · GitHub

DeepSeek-R1 stands out at reasoning tasks using a step-by-step training process, such as language, scientific thinking, and coding tasks. It features 671B overall parameters with 37B active specifications, and 128k context length.

DeepSeek-R1 constructs on the progress of earlier reasoning-focused models that improved performance by extending Chain-of-Thought (CoT) reasoning. DeepSeek-R1 takes things even more by integrating reinforcement (RL) with fine-tuning on carefully picked datasets. It developed from an earlier version, DeepSeek-R1-Zero, which relied solely on RL and revealed strong reasoning skills however had concerns like hard-to-read outputs and language inconsistencies. To resolve these restrictions, DeepSeek-R1 includes a little quantity of cold-start data and follows a refined training pipeline that mixes reasoning-oriented RL with supervised fine-tuning on curated datasets, leading to a model that attains cutting edge performance on reasoning benchmarks.

Usage Recommendations

We advise adhering to the following configurations when making use of the DeepSeek-R1 series designs, consisting of benchmarking, to achieve the expected efficiency:

– Avoid adding a system prompt; all directions should be consisted of within the user timely.
– For mathematical issues, it is advisable to consist of a regulation in your timely such as: „Please reason step by action, and put your last response within boxed .”.
– When evaluating model performance, it is suggested to perform numerous tests and balance the results.

Additional suggestions

The model’s thinking output (contained within the tags) may include more damaging content than the model’s last action. Consider how your application will utilize or display the reasoning output; you might want to reduce the reasoning output in a production setting.